CUDA_NOTES.md

Notes on CUDA

The CUDA and OS versions are selected as follows:

• CUDA: The lastest version that has image flavour devel including cuDNN available.
• OS: The latest version that has TensortRT libraries for amd64 available.
  ℹ️ It is taking quite a long time for these to be available for arm64.

Tweaks

R images

• R: Provide NVBLAS-enabled R_ and Rscript_.
  • Enabled at runtime and only if nvidia-smi and at least one GPU are present.

Environment variables

Versions

• CUDA_VERSION

Miscellaneous

• CUDA_IMAGE: The CUDA image it is derived from.
• CUDA_VISIBLE_DEVICES: If unset (default), CUDA will use all available CUDA-capable devices.
  👉 User-settable at runtime.

Settings

Default

R images

• VS Code
  • Extensions > R > Rterm: Linux: /usr/local/bin/R

Basic Linear Algebra Subprograms (BLAS)

The R images use OpenBLAS by default.

To have R and Rscript use NVBLAS instead,

1. copy the NVBLAS-enabled executables to ~/.local/bin

   for file in $(which {R,Rscript}); do
     cp "$file"_ "~/.local/bin/$(basename "$file")";
   done
   

2. set Extensions > R > Rterm > Linux: /home/USER/.local/bin/R in VS Code settings
   👉 Substitute USER with your user name.

and restart the R terminal.

ℹ️ The xgboost package benefits greatly from NVBLAS, if it is installed correctly.